This could be fun..

[Steven Nelson]

AiPi Lite

There are some of us out there that just have need to talk to machines and Toy and have them interact in our world. 

Personally I have been adding voice recognition to many of my projects for many years now. I have used several different voice recognition systems some hardware based some software based and recently through Android Apps I create using MIT's App Inventor2 programing enviroment talking to an Arduino using a HC-05 Bluetooth module and RX/TX serial data. 

Well lately I have been playing with AI Chatbots and have been wanting them to react and control things in my Real World. I have that working using the Amazon Dot and Many WiFi modules controlling Smart AC Sockets or WiFi

controlled Relays on my robot TonyBot. 

Of course doing this on my own has been difficult and an on going learning exsperiance filled with Trial and Terror as usual. 

Well apparently some folks.have been working som AI intefaces to make the Makers tasks easier. 

Here is one I just go a link to it looks interesting.

AiPi Lite...

https://aipi.com/pages/maker?fbclid=Iwb21leATVAEpwZG9mBWV4dG4DYWVtATAAYWRpZAGrL7refHI0c3J0YwZhcHBfaWQMMzUwNjg1NTMxNzI4AAEebgSMj5-YmLllZLwpz2gx915LC6iv1Wh5ZM8PoASz5UgAF0kyIaBydUmvspk_aem_8ZwcRWVFZwxWjx8hiAu7bg&utm_source=facebook&utm_medium=paid&campaign_id=120238691095850292&ad_id=120242294393310292&utm_id=120238691095850292&utm_content=120242294393310292&utm_term=120242294083170292&utm_campaign=120238691095850292

Inspection time: The period of time required to inspect a recently welded piece of metal using your bare hand.
The hotter the metal the shorter the period of inspection time

[Charles Nolan]

G'Day,

I haven't looked recently, but IBM filed many early patents on voice recognition, which might be place to start learning.

Regards,

Charles Nolan

--
You received this message because you are subscribed to the Google Groups "HomeBrew Robotics Club" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hbrobotics+...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/CAOSZ7XQ1hv65Kc%3D%2BoJJXJartfbH8nCdbSi8VGZLKfQMQKHyPZg%40mail.gmail.com.

[Chris Albertson]

No.  You absolutely do not want to go back to even 4 or 5 years of voice technology.     The technology has moved so much of late and is not expensive.   Apple has decent voice built into my watch.  How much computer power and RAM could be in a wristwatch?     The Echo Dot does voice as a side feature while it plays music and acts as a Thread router all on one tiny $2 chip.

I think the best Open Source and easy voice assistant project is built around “Wyoming Protocol”.   This is a bit of “glue” that binds together other voice projects.  So the system is modular.  You can run parts of it on different computers or even parts on the cloud.

There is more to it than you might think.   For example, you need to run a “Wake Word” detector on “hopefully” some very low-power device that listens for “Hey Robot” and then opens a connection to a more sophisticated voice processor and then sends the recognized text to so AI (cloud or local) and then the reply needs to be sent to a text-to-speech and the audio gets routed to the same location as the microphone.    Wyoming does all this but lets you pick the parts.

So for example, you could run a mon LLM on a small PC but put the microphone on a mobile robot.

This project is well supported and even has some paid staff maintaining it.  So it stays working and will be maintained.

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/828292411.1932519.1785214532375%40mail.yahoo.com.

[Charles Nolan]

G'Day Chris,

Thank you for taking the time to share your thoughts.

Best Wishes,

Charles Nolan

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/F11F7D6F-575B-4FD1-AA1E-3D7881A29DF1%40gmail.com.

[Stephen Williams]

In an Echo, and probably in your watch, only the wake word processing happens locally.  Everything else is a simple audio codec in each direction to the cloud.  All of the voice / TTS processing happens on the server where they have lots of memory and real horsepower.

That's why there are so many little, cheap "AI" devices: You just need an ESP32 with a microphone, speaker, and network connection.

We just had a presentation noting the open source wake-word detector library that runs locally.

If you need to run disconnected, or try to cut latency if you have fast enough compute, local processing is required.

Stephen

To view this discussion visit https://groups.google.com/d/msgid/hbrobotics/F11F7D6F-575B-4FD1-AA1E-3D7881A29DF1%40gmail.com.

--

Stephen D. Williams Founder: VolksDroid, Blue Scholar Foundation 650-450-8649 | fax:703-995-0407 | s...@lg.net | https://VolksDroid.org | https://BlueScholar.org | https://sdw.st/in

[Chris Albertson]

I tried building a voice system, and the software is very easy to set up.  The voice-to-text system linked below is simply “download and run,” and it runs on anything that can run Python.  And the model is very small, in the “low millions,” so we are talking about megabytes of RAM, not gigabytes.

But what is hard is the hardware.    Good microphones and good speakers are not cheap.   Also, there is the part about routing.   If you have multiple microphones, the system has to pick one and then route the response back to the right speakers.   I found it better to just use Apple’s ecosystem.    Googles would work too.   I tested with a wired heated/microphone.  Very inconvenient to use.

The killer problem was that I own an Apple Watch, and it’s clear I would never be able to build voice hardware as good as the watch. Then it is more convenient than a watch, and even if I did, I would not want to wear two watches.

This is the one case where software is easier than the hardware.

I posted a link to Wyoming before.  That is the glue that routes data between servers and local processing for wake word detection.   This link goes to the audio-to-text component.  It works very well.  But note it is only speech-to-text.  You still need to process the text.   But it will also translate.    I get English text but can speak in any language and see the English translation.   (Only able to test a few languages: Japanese, Mandarin, Spanish.  It’s perfect for robots as it can run even on a low-end Pi.)

 https://github.com/openai/whisper

On Jul 31, 2026, at 11:30 PM, Stephen Williams <s...@lig.net> wrote:

In an Echo, and probably in your watch, only the wake word processing happens locally.  Everything else is a simple audio codec in each direction to the cloud.  All of the voice / TTS processing happens on the server where they have lots of memory and real horsepower.

That's why there are so many little, cheap "AI" devices: You just need an ESP32 with a microphone, speaker, and network connection.

We just had a presentation noting the open source wake-word detector library that runs locally.

If you need to run disconnected, or try to cut latency if you have fast enough compute, local processing is required.

Stephen

On 7/28/26 9:50 AM, Chris Albertson wrote:

No.  You absolutely do not want to go back to even 4 or 5 years of voice technology.     The technology has moved so much of late and is not expensive.   Apple has decent voice built into my watch.  How much computer power and RAM could be in a wristwatch?     The Echo Dot does voice as a side feature while it plays music and acts as a Thread router all on one tiny $2 chip.

I think the best Open Source and easy voice assistant project is built around “Wyoming Protocol”.   This is a bit of “glue” that binds together other voice projects.  So the system is modular.  You can run parts of it on different computers or even parts on the cloud.

There is more to it than you might think.   For example, you need to run a “Wake Word” detector on “hopefully” some very low-power device that listens for “Hey Robot” and then opens a connection to a more sophisticated voice processor and then sends the recognized text to so AI (cloud or local) and then the reply needs to be sent to a text-to-speech and the audio gets routed to the same location as the microphone.    Wyoming does all this but lets you pick the parts.

So for example, you could run a mon LLM on a small PC but put the microphone on a mobile robot.

This project is well supported and even has some paid staff maintaining it.  So it stays working and will be maintained.

[Chris Albertson]

As it turns out, Apple Watch does the voice-to-text processing locally, on the watch.   Some of the text understanding is done locally too.   Things like “Set a timer for 5 minutes” is all done on the watch.

This started with the Series 4 hardware maybe 8 years ago.      

Apple watch has a 1.8 Ghz dual-core 64-bit ARM CPU (It runs a BSD UNIX derived kernel) and, in addtion a four-core neural processor unit and 64GB Storage, 1GB RAM. (And WiFi and BT radios plus GPS and the LCD)   Each of the four NPUs can do about 2TOPS int-8 and is optimized for matrix multiplication.    I’m not surprised the watch can do local voice commands.        You can do quite a lot with 8TOPS.

It is surprising that this watch retails for $250.   

On Jul 31, 2026, at 11:30 PM, Stephen Williams <s...@lig.net> wrote:

[Stephen Williams]

Thanks, good information!

I wonder how much you can develop for those hardware units, like the NPU?  I will try to find out at some point.

Stephen

[Chris Albertson]

You don’t develop for those NPUs. you write to Apple’s “CoreML” or “CoreAI” and then the compiler creates code that runs on whatever hardware is best at run time.   But don’t do that

Both PyTorch and TensorFlow have been ported to CoreAI and CoreML so if you use either of those, you are using whatever hardware is available on the computer/phone/watch.  It works on more than just Apple hardware